Changelog#

[1.0.0]#

Added#

  • Advanced Experimental Designer (AdvExpDesigner) — a new design engine that drops into a Campaign as designer=:

    • Continuous factors with configurable spacing (linear/geometric/logarithmic/explicit levels) and non-uniform sampling bias, drawn via stratified LHS

    • Conditional subparameters — hierarchical/nested factors (e.g. buffer type → pH range) with per-level frequency weights

    • D-optimal design optimization with parallel candidate search, seven quality metrics (D-/A-optimality, condition number, space-filling, continuous/categorical/mixed correlation), and category-correlation minimization for orthogonal categoricals

    • extend_design() for design augmentation / sequential DOE

    • Built-in visualization: histograms, mixed-correlation heatmap, quality-evolution bars, and PCA / MDS / UMAP embeddings

  • Characterization tasks — a new campaign task type (task="characterization") for level-set estimation: mapping which regions of the parameter space pass a threshold, rather than maximizing a response:

    • threshold attribute on Target; task-aware acquisition defaults on Campaign/BayesianOptimizer

    • Boundary-seeking “straddle” acquisitions: STR, RANDSTR (single-target) and MSTR, JAREX (multi-target)

    • CharacterizationEvaluator — posterior pass/fail classification, a 4-level confidence ladder (mean/70%/95% CI), joint multi-target metrics, sample-size planning, ground-truth scoring (confusion matrix / Jaccard), and largest-feasible-hypercube search

    • plot_2d_response_map (pass/fail, confidence, and continuous modes)

  • Unified RNG control (RNGManager) — coordinates numpy, torch, and Python random behind a single seed, serialized into campaign/optimizer state so a saved-and-loaded campaign reproduces exactly. One RNGManager can be shared across a campaign, optimizer, and designer; obsidian.USE_OLD_RNG_CONTROL restores legacy global-seeding behavior.

  • Custom acquisition function registry — acquisition functions are now defined as rich registry entries (implementation + hyperparameter parser + modalities + task types + constraint support). New acquisition_function_register(...) lets users plug in their own; RandomSampling (RS) is now a real implementation.

  • Classical response-surface & n-level DOE — central_composite_DOE (CCD, with rotatable/faced/custom alpha and unit-cube inscribing) and factorial_DOE_n_level (3+ level factorials); CCD exposed via ExpDesigner.initialize(method="CCD").

  • Restartable designers — both ExpDesigner and AdvExpDesigner gain save_state()/load_state(), persisted with the campaign.

  • Multi-response plotting: response_id/response_ids on parity/factor/surface/progress plots, reference-point overlay and custom objective on factor_plot, confidence bands and raw-data overlay on surface_plot, and X_suggest overlay on optim_progress.

  • Support for Python 3.10–3.12.

Modified#

  • More robust surrogate fitting — multi-restart GP fitting with best-of-N selection is now the default (max_attempts=5); fit() exposes optimizer/multi-start controls and prior-sampling initialization.

  • Optimizers can be saved/loaded unfitted (config + RNG only); save_state no longer requires a fitted model.

  • suggest() gains optim_options (passed through to BoTorch/scipy) and per-call manual_seed; fit() gains fit_options; maximize() accepts an acquisition argument; Random Search now respects fixed discrete features.

  • evaluate() now emits a standardized-prediction column per target.

Fixed#

  • Logit transform — fit now runs only on training data (not on every call, which broke for single-row inputs); degenerate/constant/all-NaN data is handled gracefully instead of producing NaN/inf, out-of-[0,1] inputs raise a clear error, and unfitted scalers raise UnfitError.

Removed#

  • Dash web app (obsidian/dash, app.py) and the [app] extra — Obsidian is now a library package.

[0.8.6]#

Added#

  • Improved methods for fitting PyTorch surrogates, including auto-stopping by parameter value norm

Modified#

  • Greatly reduced the number of samples for DNN posterior, speeding up optimization

  • Stabilized the mean estimate of ensemble surrogates by avoiding resampling

  • Disabled root caching for ensemble surrogates during optimization

  • Increased the maximum length of a category name to 32 characters

  • Bug fix for incorrect symmetry in correlation calculation of calc_ofat_ranges

  • OFAT range calcs and plots now respect minimum targets and not just maximum

[0.8.5]#

Added#

  • More optional outputs for verbose settings

  • Parameters in ParamSpace can also be indexed by name

  • Parameters now have search_space property, to modify the optimizer search space from the full space

  • Continuous parameters have search_min/search_max; Discrete parameters have search_categories

  • Constraints are now defined by Constraint class

  • Input constraints can now be included in ParamSpace, and serialized from there

  • Output constraints can now be included in Campaign, and serialized from there

  • New interface class IParamSpace to address circular import issues between ParamSpace and Constraint

Modified#

  • Optimizer and Campaign X_space attributes are now assigned using setter

  • Optimizer.maximize() appropriately recognizes fixed_var argument

Removed#

  • Torch device references and options (GPU compatibility may be re-added)

[0.8.4]#

Added#

  • Campaign X_best method

  • Optimizer X_best_f attribute(s)

  • Sequence of colors “color_list” to branding

  • Informative hoverdata for MDS plot

  • Created Product_Objective and Divide_Objective

Modified#

  • Switched all usages of X_ref = X_space.mean() to optimizer.X_best_f

  • Refactored mpl “visualize_inputs” as plotly “visualize_inputs” for better interactivity

  • Text formatting for some plotly hoverdata

[0.8.3]#

Added#

  • Default values for NParEGO scalarization_weights

  • SHAP PDP ICE plots now work with categorical values

  • Added scikit-learn to dependencies, for MDS

  • Added MDS plot

Modified#

  • SHAP PDP ICE plots must now have color and x-axis indices that are distinct

[0.8.2]#

Added#

  • Project metadata properly captured on PyPI based on changes in pyproject.toml

[0.8.1]#

Modified#

  • Fixed infobar on dash app

  • Better handling of X_space on dash app

  • Bug fixes for optim_progress

  • Improved color and axes of parity_plot

[0.8.0]#

Added#

  • Major improvements to testing and numerous small bug fixes to improve code robustness

  • Code coverage > 90%

  • New method for asserting equivalence of state_dicts during serialization

Modified#

  • Objective PyTests separated

  • Constraint PyTests separated

[0.7.13]#

Added#

  • Campaign.Explainer now added to PyTests

  • Docstrings and typing to Explainer methods

  • Campaign.out property to dynamically capture measured responses “y” or objectives as appropriate

  • Campaign.evaluate method to map optimizer.evaluate method

  • DNN to PyTests

Modified#

  • Fixed SHAP explainer analysis and visualization functions

  • Changed SHAP visualization colors to use obsidian branding

  • Moved sensitivity method from campaign.analysis to campaign.explainer

  • Moved Explainer testing from optimizer pytests to campaign pytests

  • Generalized plotting function MOO_results and renamed optim_progress

  • Campaign analysis and plotting methods fixed for multi-response campaigns

  • Greatly increased the number of samples used for DNNPosterior, increasing the stability of posterior predictions

Removed#

  • Removed code chunks regarding unused optional inputs to PDP ICE function imported from SHAP GitHub

[0.7.12]#

Added#

  • More informative docstrings for optimizer.bayesian, optimizer.predict, to explain choices of surrogate models and aq_funcs

Modified#

  • Renamed aq_func hyperparameter “Xi_f” to “inflate”

  • Moved default aq_func choices for single/multi into aq_defaults of acquisition.config

  • Fixed and improved campaign analysis methods

[0.7.11]#

Added#

  • Documentation improvements

Modified#

  • Naming conventions for modules (config, utils, base)

  • Import order convention for modules

[0.7.10]#

Added#

  • First (working) release on PyPI

[0.7.6]#

Added#

  • Added DNN surrogate model using EnsemblePosterior. Requires PosteriorList and IndexSampler during optimization

  • Added EHVI and NIPV aq_funcs

  • Added discarding of NaN X value and Y values

  • Target transforms now ignore NaN values

  • Quantile prediction to optimizer.predict() and surrogate.predict() to better suit non-GP models and non-normal distributions

Modified#

  • Generalized surrogate_botorch fitting for models which are not GPs

  • Generalized torch.dtype using global variable for import obsidian.utils.TORCH_DTYPE

  • Improved some tensor.repeat() using tensor.repeat_interleave()

  • Switched EI, NEI, and qNEHVI to Log-EI versions based on BoTorch warning

  • Dropped “q” from qNEHVI / qNParEGO names

  • Simplified surrogate model loading in BO_optimizer

  • Changed name of surrogate.args/kwargs to surrogate.hps

  • Corrected typing in target transforms

Removed#

  • Thompson sampling aq_func

[0.7.5]#

Modified#

  • Removed explainer from campaign attributes

  • Updated “features objectives” to operate on real space instead of scaled space; virtually no speed difference and tensor grads not needed

  • Improvements to campaign object and explainer object features

  • Switched all GenericMC and GenericMOMC objectives to custom objectives to simplify SOO vs MOO, and enable serialization

  • Added set_objective to campaign, and objective serialization

  • Bug fix for surrogate.save_state() with train_Y

  • Marked GPflat + categorical space test as an expected fail in pytest

  • Minor bug fixes and test improvements to improve coverage

[0.7.4]#

Modified#

  • Bug fix for qMean, made sure objective is used

  • Set up basic validation tests for parameters package

[0.7.3]#

Modified#

  • Moved plotting dependencies to main build

  • Updated Contributing section

  • Minor modifications to main Readme

  • Moved jupyterlab from docs to dev group

Removed#

  • docs/readme.md; moved contents to CONTRIBUTING

[0.7.2]#

Added#

  • Added imported members to sphinx-autodoc results using all at module level

  • Included readme on main docs page, removed redundant link

Modified#

  • Various cosmetic changes to autodoc and autosummary options

  • Shortened object names on TOC tree for readability

  • Removed autosummary from package and subpackage-level autodoc. Automated only at module level with templates

  • Split subpackages into regular and “shallow” templates, to expose different meaningful levels on the toctree

  • Updated sphinx theme with more preferred navigation features

Removed#

  • Unused dependencies from pyproject.toml

[0.7.1]#

Modified#

  • Evaluate now does not calculate aq_func by default, to speed up evaluation of objective

  • Bug fixes for optimizer.evaluate() under fringe tests

  • Updated license to GPLv3 on Readme

  • Removed Merck references and updated branding

[0.7.0]#

Added#

  • optimizer.evaluate() to handle y_predict, f_predict, o_predict, and a_predict independently of optimize.suggest()

Modified#

  • Regardless of q_batch, evaluate aq functions on each sample individually, then as a joint sample

  • Made sure that X_baseline accepted X_pending

  • Made sure that ref_point and f_best in aq_hps are now based on objectives, not raw responses

Removed#

  • Removed o_dim and optim_type from optimizer.attrs to better support stateless operation

  • No longer calculate hypervolume or pareto front on raw responses; only after considering objectives

  • Removed y_pareto from optimizer.attrs

  • Temporarily removed campaign._analyze due to bugs with_hv

[0.6.11]#

Added#

  • Added plotting module to test coverage

  • Reloading/fitting f_train in optimizer.load_state

  • Added default_campaign.json to test directory, for faster testing using pre-fit objects

  • Added matplotlib plotting library

  • Can now provide X_pending to suggest, so that users can manually do iterative suggestions

Modified#

  • Various plotting bug fixes

  • Ensure that paramspace encode/unit_map return double dtypes

[0.6.10]#

Added#

  • Added getitem to ParamSpace to enable X_space[idx]

Modified#

  • Overhaul of all encode/decode and map/demap functions using a decorator to handle robust typing

  • Bug fix for SHAP_explain resulting from object dtypes with new encode functions

[0.6.9]#

Added#

  • Added notebook tutorials to docs

  • repr method for Target

Modified#

  • Bug fixes in factor_plot with X_ref provided

  • Fixed bug where categorical encoding wouldn’t work if all categories were numerical strings

  • Changed categorical OH-encode separator from _to ^ and protected against usage in X_names

[0.6.8]#

Modified#

  • Dash app updates with module refactoring

  • Minor refactoring of param_space discrete handling

  • Replaced assert statements with appropriate Python base exceptions, added custom obsidian exceptions as necessary

  • Moved benchmark module to obsidian.experiment

Removed#

  • Removed deprecated examples

[0.6.7]#

Added#

  • Dev capabilities for document generation using sphinx

  • Module docstrings

[0.6.6]#

Modified#

  • Overhaul of class and method docstrings

Removed#

  • Removed parameter.type and replaced with isinstance checking and class.name loading

[0.6.5]#

Added#

  • Custom exception handling

  • Composite objectives

  • Utopian point support for SOO and MOO

  • Bounded target support for SOO and MOO

  • Bounding for only selected targets in multi-output scenarios

Modified#

  • Completed refactored weighted MOO to separate out component parts of scalarization, utopian distance, norming, and bounding

  • Greatly simplified the structure of aq_kwargs based on the above

Removed#

  • weightedMOO acquisition function

  • Dynamic reference point utility

[0.6.4]#

Added#

  • More custom objectives

Modified#

  • Switched default single-objective aq from EI to NEI

Removed#

  • Simplex sampling of weights for weightedMOO; if weights aren’t provided, even weights are used

[0.6.3]#

Modified#

  • Moved “explain” functionality to base optimizer

  • Added backup exceptions to catch fit_gpytorch

  • Enabled multi-output objectives to expand single-response models (e.g. using X features)

  • Enabled single-output objectives to condense multi-response models (e.g. scalarization)

  • Updated demo notebooks

  • Fixed error with calling campaign._profile_max after every fit

Removed#

  • Redundant objective formulations

  • Removed GPFlex model, which is now redundant with multi-output objective

[0.6.2]#

Modified#

  • Updated all plotly plotting methods for new optimizer methods

  • Updated typing and enabled multi-objective on custom aq functions

[0.6.1]#

Modified#

  • Added new features to campaign object

  • Moved “seed” kwarg to ExpDesigner.init, consistent with BayesianOptimizer

  • Added hypervolume and pareto calculations (incl. pf_distance) to base optimizer

  • Fixed bug with target_transforms sharing hyperparameters because of bad initialization

[0.6.0]#

Added#

  • Task parameters and multi-task learning

  • Added “index” objective to select tasks for optimization in multi-task learning

  • Fixed bugs related to torch dtype mismatches

[0.5.7]#

Modified#

  • Fixed cat_dims specification for surrogate models so that they do not include ordinal params

  • Overhauled f_transform approach to avoid scikit-learn and be more customizable

  • Enforced abstractmethods on various classes

[0.5.7]#

Modified#

  • Fixed cat_dims specification for surrogate models so that they do not include ordinal params

  • Overhauled f_transform approach to avoid scikit-learn and be more customizable

  • Enforced abstractmethods on various classes

[0.5.6]#

Modified#

  • Added utopian point subtraction to all scalarization methods, made optional also

  • Moved PI_bounded weightedMOO to its own objective called “boundedMOO”

  • Fixed error where default hyperparameters were being written back to objects outside of the optimizer

  • Fixed bugs with new parameter types in experiment design modules

  • Added pytest parametrization to improve scope of tests

  • Added pytest attributes (slow, fast) to manage speed

[0.5.5]#

Modified#

  • Fixed bug with _fixed_features generation when fixed_var is specified

  • Updated all MOO custom aq functions to match current BoTorch patterns

Removed#

  • PI_bounded acquisition function, due to various issues. MOO_weighted with PI_bounded+weights scalarization does work

[0.5.4]#

Modified#

  • Added Param_Discrete_Numeric class (parent Param_Discrete)

  • Added Param_Observational subclass to parent Continuous class which can be used for fitting but avoid optimization

[0.5.3]#

Modified#

  • Implemented checks and validation to enforce the order of X, y, and targets across ParamSpace, Optimizer, Surrogate as appropriate

    • Note: Only Optimizer can handle extraneous or re-ordered columns, but they will be processed before passing to Surrogate

[0.5.2]#

Modified#

  • Fixed input/output constraints

  • Added de/transformation to output constraints (applied to target samples)

  • Added de/transformation to input constraints (applied to coefficients and RHS)

  • Implemented a de/transform map for ParamSpace in order to handle constraints in the encoded input space

  • Implemented a non-linear input constraint which keeps the range of one dimension in a joint optimization < 1% of the max-min

[0.5.1]#

Added#

  • Constrained multi-objective example notebook

  • Allowed optimizer.suggest() on a subset of fit responses

  • Enabled optimizer.maximize() for multi-response models based on the above

  • Added custom multi-output objective class

  • Fixed constraints specification, as we were using output constraints. Added input constraints to TODO

  • Added parameter name to lb/ub labels in optimizer.predict() to avoid index issues

Modified#

  • Custom constraints are now specified as a constructor, so that parameters can be added and a callable is returned

[0.5.0]#

Added#

  • New object-oriented design for several classes: Campaign Parameter ParamSpace Target Objective and Constraint

On this page
Show Source